Atmospheric particulate matter source analysis method for constructing pollution matrix based on convolutional neural network

Abstract

The invention discloses an atmospheric particulate source analysis method for constructing a pollution matrix based on a convolutional neural network, and the method comprises the steps: employing SO2 concentration, NO2 concentration, CO concentration, PM2.5 concentration, PM10 concentration and humidity as pollution index parameters, constructing the data of each pollution index parameter of a known pollution source into basic feature matrix portraits of different source types; and after being learned by the image convolutional neural network model, the image is used for analyzing the feature matrix image of the unknown pollution source. According to the invention, the atmospheric particulate matter source analysis model after learning optimization is utilized to analyze the feature matrix picture of the unknown pollution source, so that the pollution type of each sample in the unknown pollution source can be timely and accurately obtained; moreover, pollution source change process diagrams corresponding to different pollution periods of the unknown pollution source in the pollution period can be obtained, an efficient source judgment method can be provided for an environment law enforcement department, and the method is low in use cost, high in timeliness and good in application prospect.

Description

Technical Field

[0001] The invention belongs to the field of environmental pollution monitoring and relates to an atmospheric particulate matter source analysis method based on building a pollution matrix using a convolutional neural network. Background Art

[0002] Atmospheric particulate matter source analysis technology is a technology that qualitatively or quantitatively identifies the source of atmospheric particulate matter pollution through chemical, physical, mathematical and other methods. Source analysis has a high degree of recognition in environmental monitoring. At present, the source analysis methods in environmental monitoring mainly include manual source analysis, mass spectrometry (TOF-MS) and single data convolutional neural network atmospheric particulate matter source analysis method. Among them, manual source analysis and mass spectrometry have the defects of high cost, difficulty in covering high-intensity monitoring periods, and inability to be used on a large scale. The single data convolutional neural network atmospheric particulate matter source analysis method has the following defects: (a) large amount of data processing and poor analysis timeliness; (b) there are defects such as overfitting or induced erroneous results by high concentration values, which makes it difficult to achieve widespread use. Therefore, it is urgent to develop a source analysis method with low cost, wide equipment coverage, accurate data analysis and good operation effect, which is of great practical significance for the prevention and control of air pollution. Summary of the invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for analyzing the sources of atmospheric particulate matter based on a convolutional neural network to construct a pollution matrix, which has low cost, strong versatility, large data processing volume, good real-time data analysis, high precision and good source reliability.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions: A method for atmospheric particulate matter source analysis based on a convolutional neural network to construct a pollution matrix includes the following steps: Step S1, obtaining the data of pollution index parameters in different samples of known pollution sources during the pollution period; the pollution index parameters are SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10 concentration and humidity; Step S2, obtaining basic feature matrix images of different source categories according to the data of each pollution index parameter in different samples of the known pollution source within the pollution period obtained in step S1; Step S3, combining the basic feature matrix images of different source categories in step S2 into a picture sample set, using an image convolutional neural network model for learning, and obtaining an atmospheric particulate matter source analysis model; Step S4, collecting data of various pollution index parameters in different samples of the unknown pollution source during the pollution period, and obtaining characteristic matrix images corresponding to different samples in the unknown pollution source; Step S5: input the characteristic matrix images corresponding to different samples of the unknown pollution source obtained in step S4 into the atmospheric particulate matter source analysis model obtained in step S3 for analysis to obtain the pollution types of different samples of the unknown pollution source.

[0005] The above-mentioned atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix is ​​further improved. In step S2, the method for establishing the basic feature matrix image of different source categories includes the following steps: Step S2-1: SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10 Concentration and humidity were fitted into a 6 × 6 matrix as pollution index parameters and used as a blank matrix; Step S2-2, normalizing the data of each pollution index parameter in different samples of the known pollution source within the pollution period obtained in step S1 to obtain characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution source; Step S2-3, inputting the characteristic parameters corresponding to each pollution index parameter in different samples of known pollution sources into a blank matrix, and obtaining a basic characteristic matrix image of the characteristic parameters corresponding to different samples of known pollution sources; Step S2-4: According to the discriminant between the set different characteristic parameters and the source category, the basic characteristic matrix images of the characteristic parameters corresponding to different samples of the known pollution source are subjected to source category determination to obtain basic characteristic matrix images of different source categories.

[0006] The above-mentioned atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix is ​​further improved. In step S2-2, the Z-score normalization method is used to normalize the data of each pollution index parameter in different samples of the known pollution source within the pollution period obtained in step S1, and the characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution source are recorded as 0 or 1, so as to obtain the characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution source.

[0007] The above-mentioned atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix is ​​further improved. In step S2-2, the normalization process also includes: weighting SO2 concentration, NO2 concentration, and CO concentration; the SO2 concentration is greater than 20μg / m 3 When the NO2 concentration is greater than 70 μg / m 3 When the CO concentration is greater than 1 mg / m 3 When , the corresponding characteristic parameter is recorded as 1, otherwise it is not recorded.

[0008] The above-mentioned atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix is ​​further improved. In step S2-4, the discriminant between the set different characteristic parameters and source categories includes a mapping relationship between the different characteristic parameters and the source categories; the mapping relationship between the different characteristic parameters and the source categories in the discriminant between the set characteristic parameters and different source categories is: S, N, C, D, F, and R represent SO2 concentration, NO2 concentration, CO concentration, PM concentration, etc. 2.5 Concentration, PM 10 Characteristic parameters of concentration and humidity, when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥PM 2.5 When the actual concentration is ×0.8, the source category of the corresponding basic feature matrix image is dust source; when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When S=0, N=0, C=0, D=0, F=0, R=0, the source category of the corresponding basic characteristic matrix image is secondary inorganic source; when S=1, N=0, C=1, D=1, F=1, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When S=0, N=1, C=0, D=0, F=0, R=0, the source category of the corresponding basic feature matrix image is the motor vehicle exhaust source; when S=1, N=0, C=0, D=0, F=0, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When , the source category of the corresponding basic feature matrix portrait is industrial source; in other cases, the source category of the corresponding basic feature matrix portrait is other source.

[0009] The above-mentioned atmospheric particulate matter source analysis method for constructing a pollution matrix based on a convolutional neural network is further improved. In step S3, the atmospheric particulate matter source analysis model includes a mapping relationship between different characteristic parameters and source categories; the mapping relationship between different characteristic parameters and source categories in the atmospheric particulate matter source analysis model is: S, N, C, D, F, and R represent SO2 concentration, NO2 concentration, CO concentration, PM concentration, etc. 2.5 Concentration, PM 10 Characteristic parameters of concentration and humidity, when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m3 When S=1, N=0, C=1, D=0, F=0, R=0, PM 10 Actual concentration>80μg / m 3 , PM 2.5 Actual concentration ≥100 μg / m 3 When S=0, N=0, C=0, D=0, F=0, R=0, the source category of the corresponding characteristic matrix image is determined to be biased towards coal-fired sources; when S=0, N=0, C=0, D=0, F=0, R=0, the source category of the corresponding characteristic matrix image is determined to be biased towards secondary inorganic sources; when S=1, N=0, C=1, D=1, F=1, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When S=0, N=1, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When S=1, N=0, C=0, D=0, F=0, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When the source category of the corresponding feature matrix image is determined to be industrial source, in other cases, the source category of the corresponding feature matrix image is determined to be other source.

[0010] The above-mentioned atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix is ​​further improved. In step S4, the method for establishing the characteristic matrix image corresponding to different samples in the unknown pollution source includes the following steps: Step S4-1: SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10 Concentration and humidity were fitted into a 6 × 6 matrix as pollution index parameters and used as a blank matrix; Step S4-2: normalize the data of each pollution index parameter in different samples of the unknown pollution source during the pollution period to obtain characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution source.

[0011] Step S4-3: input the characteristic parameters corresponding to each pollution index parameter in different samples of known pollution sources into the blank matrix to obtain the characteristic matrix images corresponding to different samples of unknown pollution sources.

[0012] The above-mentioned atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix is ​​further improved, and step S5 also includes the following processing: using the autoregressive model to determine the pollution period of the characteristic matrix image corresponding to different samples in the unknown pollution source, and drawing the pollution source change effect diagram corresponding to the different pollution periods of the unknown pollution source within the pollution period.

[0013] The above-mentioned atmospheric particulate matter source analysis method based on the convolutional neural network to construct the pollution matrix is ​​further improved based on the PM 2.5 The concentration determines the pollution period of the characteristic matrix images corresponding to different samples in the unknown pollution source, as follows: 2.5 Concentration <30 μg / m 3 When the characteristic matrix image of the corresponding sample is judged as the matrix of the non-polluting period; when 31μg / m 3 <PM 2.5 Concentration <70 μg / m 3 , the characteristic matrix image of the corresponding sample is determined as the matrix of the pollution coming period; when 71μg / m 3 <PM 2.5 Concentration < 500 μg / m 3 When PM 2.5 The concentration decreased for three consecutive values ​​and then recovered to PM 2.5 Concentration <30 μg / m 3 Before, the characteristic matrix image of the corresponding sample is determined to be the pollution decay period matrix.

[0014] Compared with the prior art, the advantages of the present invention are: (1) In view of the shortcomings of existing source analysis methods, such as high cost, difficulty in covering high-intensity monitoring periods, inability to be used over a large area, poor analysis timeliness, and low accuracy, the present invention creatively proposes a method for atmospheric particulate matter source analysis based on a convolutional neural network to construct a pollution matrix, using SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10Concentration and humidity are pollution index parameters, and the data of each pollution index parameter in different samples of known pollution sources during the pollution period are constructed into basic feature matrix images of different source categories. By fitting the data of the six parameters to the basic feature matrix image and using the image convolutional neural network model for learning, on the one hand, the data processing volume can be increased by fitting, and at the same time, the output of the image convolutional neural network model can be reduced. Compared with the single data convolutional neural network model, the output can be reduced by 6 times, avoiding the risk of overfitting or high numerical induction of erroneous results. On the other hand, the image convolutional neural network model is used to train the basic feature matrix images of different source categories, which can effectively avoid the interference of secondary inorganic sources on the data ratio, which is conducive to reducing the source parsing analysis error, and the source parsing analysis error is less than 3%. More importantly, When using the image convolutional neural network model for learning, there is no need to manually design a feature extractor, burst values ​​will not induce model learning errors, and a single image segmentation task can be used. In addition, the image can be ensured to have higher accuracy through multi-layer convolution and pooling, avoiding misjudgment caused by high single data, thereby ensuring the long-term accuracy of the data. On this basis, the data of each pollution index parameter in different samples of the unknown pollution source during the pollution period are constructed into a feature matrix image, and analyzed using the atmospheric particulate matter source analysis model after learning optimization. Not only can the pollution type of each sample in the unknown pollution source be obtained in a timely and accurate manner, but also the pollution source change process diagram corresponding to different pollution periods of the unknown pollution source during the pollution period can be obtained. This can provide environmental law enforcement departments with an efficient source identification method with low cost, strong timeliness and good application prospects.

[0015] (2) In the present invention, the starting point is no longer from the industry source, but from the reaction process in the atmosphere, which can better explain the situation of pollutants entering the atmosphere at different pollution periods.

[0016] (3) Compared with the manual source analysis method, the present invention has lower cost and the total cost is reduced by 60%.

[0017] (4) In the present invention, the data of pollution index parameters in different samples of known pollution sources during the pollution period can be collected from national control stations, provincial control stations or six-parameter mobile vehicles. It has strong versatility, is easy to collect, and has a relatively low collection cost.

[0018] (5) In the present invention, the accuracy of rapid analysis of high pollution periods (365 days) and urban air quality is comprehensively improved (the error is less than 3% compared with the manual source analysis method). BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0020] Figure 1 Schematic diagram of the process flow of the atmospheric particulate matter source analysis method for constructing a pollution matrix based on a convolutional neural network in Example 1 of the present invention.

[0021] Figure 2 This is a basic feature matrix image corresponding to the industrial source in Example 1 of the present invention.

[0022] Figure 3 This is a source analysis result diagram of the atmospheric particulate matter source analysis method based on convolutional neural network to construct a pollution matrix in Example 1 of the present invention.

[0023] Figure 4 This is a source parsing result diagram of the manual source parsing method in Example 1 of the present invention.

[0024] Figure 5 This is a process diagram of the entire pollution period of the atmospheric particulate matter source analysis method based on convolutional neural network to construct a pollution matrix in Example 1 of the present invention.

[0025] Figure 6 This is a comparison chart of the implementation costs of different source analysis methods in Example 1 of the present invention.

[0026] Figure 7 This is a comparison chart of the recognition of short-term data by different source analysis methods in Example 2 of the present invention.

[0027] Figure 8 This is a comparison chart of the recognition of long-term data by different source analysis methods in Example 2 of the present invention. DETAILED DESCRIPTION

[0028] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0029] Aiming at the shortcomings of existing source analysis methods, such as high cost, difficulty in covering high-intensity monitoring periods, inability to use on a large scale, poor analysis timeliness, and low precision, this paper creatively proposes a method for atmospheric particulate matter source analysis based on a convolutional neural network to construct a pollution matrix, using SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10Concentration and humidity are pollution index parameters, and the data of each pollution index parameter in different samples of known pollution sources during the pollution period are constructed into basic feature matrix images of different source categories. By fitting the data of the six parameters to the basic feature matrix image and using the image convolutional neural network model for learning, on the one hand, the data processing volume can be increased by fitting, and at the same time, the output of the image convolutional neural network model can be reduced. Compared with the single data convolutional neural network model, the output can be reduced by 6 times, avoiding the risk of overfitting or high numerical induction of erroneous results. On the other hand, the image convolutional neural network model is used to train the basic feature matrix images of different source categories, which can effectively avoid the interference of secondary inorganic sources on the data ratio, which is conducive to reducing the source parsing analysis error, and the source parsing analysis error is less than 3%. More importantly, When using the image convolutional neural network model for learning, there is no need to manually design a feature extractor, burst values ​​will not induce model learning errors, and a single image segmentation task can be used. In addition, the image can be ensured to have higher accuracy through multi-layer convolution and pooling, avoiding misjudgment caused by high single data, thereby ensuring the long-term accuracy of the data. On this basis, the data of each pollution index parameter in different samples of the unknown pollution source during the pollution period are constructed into a feature matrix image, and analyzed using the atmospheric particulate matter source analysis model after learning optimization. Not only can the pollution type of each sample in the unknown pollution source be obtained in a timely and accurate manner, but also the pollution source change process diagram corresponding to different pollution periods of the unknown pollution source during the pollution period can be obtained. This can provide environmental law enforcement departments with an efficient source identification method with low cost, strong timeliness and good application prospects.

[0030] In order to better understand the innovation of the present invention, the present invention also provides the following embodiments.

[0031] Example 1 A method for atmospheric particulate matter source analysis based on convolutional neural network to construct pollution matrix, the process flow chart is as follows Figure 1 As shown, the following steps are included: Step S1, obtain the data of pollution index parameters of different samples of known pollution sources (industrial sources, coal-fired sources, motor vehicle exhaust sources, biomass combustion sources, dust sources and secondary inorganic sources) during the pollution period, where the pollution index parameters are SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10 concentration and humidity.

[0032] In step S1, conventional atmospheric parameter collection equipment is used to collect data on various pollution sources (motor vehicle exhaust, various industrial sources, coal-burning sources, dust sources, biomass combustion sources, etc.). The required six-parameter data are required to be data collected by national control stations, provincial control stations or six-parameter UAVs, and meet national air quality standards. At the same time, the required existing source analysis data must comply with the "Technical Guidelines for Source Analysis of Atmospheric Particulate Matter", and the corresponding pollution sources must account for more than 30% of the total in a single source analysis period.

[0033] Step S2: According to the data of each pollution index parameter in different samples of the known pollution source during the pollution period obtained in step S1, basic feature matrix images of different source categories are obtained, wherein the method for establishing basic feature matrix images of different source categories includes the following steps: Step S2-1: SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10 Concentration and humidity were fitted into a 6 × 6 matrix as pollution index parameters and used as a blank matrix.

[0034] Step S2-2, normalizing the data of each pollution index parameter in different samples of the known pollution source within the pollution period obtained in step S1 to obtain characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution source.

[0035] In step S2-2, the data of each pollution index parameter in different samples of all known pollution sources within the pollution period are imported into MATLAB, and the data of each pollution index parameter in different samples of the known pollution sources within the pollution period obtained in step S1 are normalized using the Z-score normalization method. The characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution sources are recorded as 0 or 1, and the characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution sources are obtained.

[0036] In step S2-2, all data are normalized and averaged. After processing, the normalized value is compared with the average deviation value to generate characteristic parameters and save them. If the normalized value is greater than the average deviation value, it is recorded as "1", and if the normalized value is less than or equal to the average deviation value, it is recorded as "0". At the same time, the normalization process also includes: weighting the SO2 concentration, NO2 concentration, and CO concentration, where the SO2 concentration is greater than 20μg / m 3 When the NO2 concentration is greater than 70 μg / m 3 When the CO concentration is greater than 1 mg / m 3 When , the corresponding characteristic parameter is recorded as 1, otherwise it is not recorded.

[0037] Step S2-3, input the characteristic parameters corresponding to each pollution index parameter in different samples of known pollution sources into the blank matrix to obtain a basic characteristic matrix image of the characteristic parameters corresponding to different samples of known pollution sources.

[0038] Step S2-4: According to the discriminant between the set characteristic parameters and different source categories, the basic characteristic matrix images of the characteristic parameters corresponding to different samples of the known pollution source are subjected to source category determination to obtain basic characteristic matrix images of different source categories. Figure 2 As shown, Figure 2 It is one of the basic feature matrix images corresponding to the industrial source.

[0039] In step S2-4, the discriminant between the characteristic parameters and different source categories includes the mapping relationship between the different characteristic parameters and the source categories, and the mapping relationship between the different characteristic parameters and the source categories is: S, N, C, D, F, and R represent SO2 concentration, NO2 concentration, CO concentration, PM concentration, etc. 2.5 Concentration, PM 10 Characteristic parameters of concentration and humidity, when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥PM 2.5 When the actual concentration is ×0.8, the source category of the corresponding basic feature matrix image is dust source; when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When S=0, N=0, C=0, D=0, F=0, R=0, the source category of the corresponding basic characteristic matrix image is secondary inorganic source; when S=1, N=0, C=1, D=1, F=1, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When S=0, N=1, C=0, D=0, F=0, R=0, the source category of the corresponding basic feature matrix image is the motor vehicle exhaust source; when S=1, N=0, C=0, D=0, F=0, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When , the source category of the corresponding basic feature matrix portrait is industrial source; in other cases, the source category of the corresponding basic feature matrix portrait is other source.

[0040] Step S3: Save the basic feature matrix images of different source categories in step S2 to form a sample set of pictures (a total of 200 pictures) on the local disk of the computer, use the image convolutional neural network model for learning, and record the corresponding pollution source category names to obtain the atmospheric particulate matter source analysis model.

[0041] In step S3, the atmospheric particulate matter source analysis model includes a mapping relationship between different characteristic parameters and source categories, and the mapping relationship between different characteristic parameters and source categories in the atmospheric particulate matter source analysis model is: S, N, C, D, F, and R represent SO2 concentration, NO2 concentration, CO concentration, PM concentration, and 2.5 Concentration, PM 10 Characteristic parameters of concentration and humidity, when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When S=1, N=0, C=1, D=0, F=0, R=0, PM 10 Actual concentration>80μg / m 3 , PM 2.5 Actual concentration ≥100 μg / m 3 When S=0, N=0, C=0, D=0, F=0, R=0, the source category of the corresponding characteristic matrix image is biased towards secondary inorganic sources; when S=1, N=0, C=1, D=1, F=1, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When S=0, N=1, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When S=1, N=0, C=0, D=0, F=0, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When , the source category of the corresponding feature matrix picture is industrial source; in other cases, the source category of the corresponding feature matrix picture is other source.

[0042] Step S4: collect data on various pollution index parameters in different samples of the unknown pollution source in City A during the pollution period, and obtain feature matrix images corresponding to different samples in the unknown pollution source, for a total of 720 feature matrix images.

[0043] Step S5: Input the feature matrix images corresponding to different samples of the unknown pollution source obtained in step S4 into the atmospheric particulate matter source analysis model obtained in step S3 for analysis, pass the similarity results of the final model analysis features, output the pollution source category determination results, and obtain the pollution types of different samples of the unknown pollution source. The results are as follows: Figure 3 shown.

[0044] like Figure 3 As shown, when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When S=0, N=0, C=0, D=0, F=0, R=0, it is determined to be a partial dust source, with a total of 71 pictures (a total of 720 groups of pictures, accounting for 9.8%); when S=0, N=0, C=0, D=0, F=0, R=0, it is determined to be a partial secondary inorganic source, with a total of 205 pictures (a total of 720 groups of pictures, accounting for 28.43%); when S=1, N=0, C=0, D=0, F=0, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When S=1, N=0, C=1, D=0, F=0, R=0, PM 10 Actual concentration>80μg / m 3 , PM 2.5 Actual concentration ≥100 μg / m 3 It was judged as a coal-burning source, with a total of 148 pictures (a total of 720 groups of pictures, accounting for 20.59%). When S=1, N=0, C=1, D=1, F=1, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When S=0, N=1, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When the source is motor vehicle exhaust, it is judged to be 155 pictures (a total of 720 groups of pictures, accounting for 21.57%), and other sources total 36 pictures (a total of 720 groups of pictures, accounting for 4.9%).

[0045] In order to verify the accuracy of the model, manual source analysis was performed on atmospheric particulate matter in City A. The source analysis results are as follows: Figure 4 As shown. Figure 4It can be seen that among the samples of unknown sources, dust sources accounted for 9%, secondary inorganic sources accounted for 29.2%, coal-fired sources accounted for 20%, industrial sources accounted for 11%, motor vehicle exhaust sources accounted for 22%; biomass combustion sources accounted for 4%, and other sources accounted for 4.8%.

[0046] Depend on Figure 3 and Figure 4 From the results in , it can be seen that the analysis errors of the two source analysis methods are less than 3%, indicating that the source analysis method provided by the present invention is highly accurate and reliable.

[0047] Step S5 also includes the following processing: using the autoregressive model and according to the PM in the sample 2.5 The concentration determines the pollution period of the characteristic matrix images corresponding to different samples in the unknown pollution source, specifically when PM 2.5 Concentration <30 μg / m 3 When the characteristic matrix image of the corresponding sample is judged as the matrix of the non-polluting period; when 31μg / m 3 <PM 2.5 Concentration <70 μg / m 3 , the characteristic matrix image of the corresponding sample is determined as the matrix of the pollution coming period; when 71μg / m 3 <PM 2.5 Concentration < 500 μg / m 3 When PM 2.5 The concentration decreased for three consecutive values ​​and then recovered to PM 2.5 Concentration <30 μg / m 3 Before, the characteristic matrix image of the corresponding sample is determined as the pollution decay period matrix. According to the pollution period sequence, the characteristic matrix images corresponding to different pollution periods are plotted to draw the pollution source change effect diagram corresponding to different pollution periods in the pollution period. The results are as follows Figure 5 shown.

[0048] Depend on Figure 5 It can be seen that after learning the parameter matrix of the pollution period through the image convolutional neural network model and the autoregressive model, the simulation analysis of the off-site abnormal intervals under the same background can be used. The results show that particulate matter and humidity have special contributions. The relative humidity of the site in City A affects the secondary generation of inorganic aerosols by affecting the heterogeneous reaction of the liquid phase on the surface of particulate matter, resulting in an increase in particulate matter concentration. It can be seen that compared with other types of source analysis methods, the method of the present invention better explains the problem of abnormal increase of particulate matter under the same environmental background, and the real-time pollution process analyzed by the model in the present invention is an advantage that manual source analysis and atmospheric particulate matter (single data) convolutional neural network source analysis do not have.

[0049] In actual application, compared with traditional manual source analysis, the source analysis method provided by the present invention increases the analysis time to 365 days, and the equipment cost and sampling method cost are reduced by about 60%. Figure 6 shown.

[0050] Example 2 The practical applicability of the source apportionment method of the present invention and the conventional atmospheric particulate matter (single data) convolutional neural network source apportionment method is investigated, and the operation steps include the following: The data of various pollution index parameters in different samples of the unknown pollution sources in City A during the pollution period collected in Example 1 were randomly selected to form a validation set, and the method of Example 1 of the present invention and the convolutional neural network source analysis method of atmospheric particulate matter (single data) were verified. The two methods were observed and the correlation coefficient was analyzed with the manual source analysis of City A.

[0051] Depend on Figure 7 It can be seen that in the short term (data within 30 days), the R=0.76 in the atmospheric particulate matter convolutional neural network source analysis method, while the R=0.80 in the source analysis method of the present invention has better analysis accuracy. Figure 8 As shown, in the long term (data within 180 days), the parsing accuracy of the source parsing method of the present invention is significantly better than that of the conventional atmospheric particulate matter (single data) convolutional neural network source parsing method. The reason is that the external transmission of particulate matter increases in autumn and winter, and the number of sudden outliers increases. The single data convolutional neural network will induce erroneous learning due to the sudden outlier data, resulting in a decrease in accuracy R=0.36. In contrast, in the source parsing method of the present invention, when the image convolutional neural network model is used for learning, there is no need to manually design a feature extractor, and sudden values ​​will not induce model erroneous learning. A single image segmentation task can be used, and the image can be ensured to have higher accuracy through multi-layer convolution and pooling, avoiding misjudgment caused by high single data. This can ensure the accuracy of the data over a long period of time, and the accuracy is still very high, with R=0.81.

[0052] Depend on Figure 7 and Figure 8 It can be seen that compared with the conventional atmospheric particulate matter convolutional neural network source analysis method, the method of the present invention has better accuracy.

[0053] From the above results, it can be seen that compared with the conventional source analysis method, in the method of the present invention, the data of each pollution index parameter in different samples of the unknown pollution source during the pollution period are constructed into a feature matrix image, and the atmospheric particulate matter source analysis model after learning optimization is used for analysis, which can not only timely and accurately obtain the pollution type of each sample in the unknown pollution source, but also obtain the pollution source change process diagram corresponding to the different pollution periods of the unknown pollution source during the pollution period, which can provide an efficient source identification method for environmental law enforcement departments, and has low cost, strong timeliness and good application prospects. In addition, the method of the present invention also has the following advantages: (a) It no longer starts from the source of the industry, but from the process of reaction in the atmosphere, which can better explain the situation of pollutants entering the atmosphere during different pollution periods. (b) Compared with the manual source analysis method, the cost is lower, and the total cost is reduced by 60%. (c) The data of pollution index parameters in different samples of known pollution sources during the pollution period can be collected from national control stations, provincial control stations or six-parameter rovers, which has strong versatility, is easy to collect, and has a relatively low collection cost. (d) The accuracy of rapid analysis of high-pollution periods (365 days) and urban air quality has been comprehensively improved (the error is less than 3% compared with the manual source analysis method).

[0054] The above embodiments are only preferred implementations of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for source analysis of atmospheric particulate matter based on a convolutional neural network to construct a pollution matrix, characterized in that: The following steps are involved: Step S1, obtaining the data of pollution index parameters in different samples of known pollution sources during the pollution period; the pollution index parameters are SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10 concentration and humidity; Step S2, obtaining basic feature matrix images of different source categories according to the data of each pollution index parameter in different samples of the known pollution source within the pollution period obtained in step S1; Step S3, combining the basic feature matrix images of different source categories in step S2 into a picture sample set, using an image convolutional neural network model for learning, and obtaining an atmospheric particulate matter source analysis model; Step S4, collecting data of various pollution index parameters in different samples of the unknown pollution source during the pollution period, and obtaining characteristic matrix images corresponding to different samples in the unknown pollution source; Step S5: input the characteristic matrix images corresponding to different samples of the unknown pollution source obtained in step S4 into the atmospheric particulate matter source analysis model obtained in step S3 for analysis to obtain the pollution types of different samples of the unknown pollution source.

2. The atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix according to claim 1 is characterized in that: In step S2, the method for establishing the basic feature matrix portraits of different source categories includes the following steps: Step S2-1: SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10 Concentration and humidity were fitted into a 6 × 6 matrix as pollution index parameters and used as a blank matrix; Step S2-2, normalizing the data of each pollution index parameter in different samples of the known pollution source within the pollution period obtained in step S1 to obtain characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution source; Step S2-3, inputting the characteristic parameters corresponding to each pollution index parameter in different samples of known pollution sources into a blank matrix, and obtaining a basic characteristic matrix image of the characteristic parameters corresponding to different samples of known pollution sources; Step S2-4: According to the discriminant between the set different characteristic parameters and the source category, the basic characteristic matrix images of the characteristic parameters corresponding to different samples of the known pollution source are subjected to source category determination to obtain basic characteristic matrix images of different source categories.

3. The atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix according to claim 2 is characterized in that: In step S2-2, the Z-score normalization method is used to normalize the data of each pollution index parameter in different samples of the known pollution source within the pollution period obtained in step S1, and the characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution source are recorded as 0 or 1, so as to obtain the characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution source.

4. The method for source analysis of atmospheric particulate matter based on convolutional neural network to construct pollution matrix according to claim 3 is characterized in that: In step S2-2, the normalization process also includes: weighting the SO2 concentration, NO2 concentration, and CO concentration; the SO2 concentration is greater than 20 μg / m 3 When the NO2 concentration is greater than 70 μg / m 3 When the CO concentration is greater than 1 mg / m 3 When , the corresponding characteristic parameter is recorded as 1, otherwise it is not recorded.

5. The atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix according to claim 2 is characterized in that: In step S2-4, the discriminant between the set different characteristic parameters and source categories includes a mapping relationship between the different characteristic parameters and the source categories; the mapping relationship between the different characteristic parameters and the source categories in the discriminant between the set characteristic parameters and the different source categories is: S, N, C, D, F, and R represent SO2 concentration, NO2 concentration, CO concentration, PM concentration, respectively. 2.5 Concentration, PM 10 Characteristic parameters of concentration and humidity, when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥PM 2.5 When the actual concentration is ×0.8, the source category of the corresponding basic feature matrix image is dust source; when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When S=0, N=0, C=0, D=0, F=0, R=0, the source category of the corresponding basic characteristic matrix image is secondary inorganic source; when S=1, N=0, C=1, D=1, F=1, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When S=0, N=1, C=0, D=0, F=0, R=0, the source category of the corresponding basic feature matrix image is the motor vehicle exhaust source; when S=1, N=0, C=0, D=0, F=0, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When , the source category of the corresponding basic feature matrix portrait is industrial source; in other cases, the source category of the corresponding basic feature matrix portrait is other source.

6. The atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix according to any one of claims 1 to 5, characterized in that: In step S3, the atmospheric particulate matter source analysis model includes a mapping relationship between different characteristic parameters and source categories; the mapping relationship between different characteristic parameters and source categories in the atmospheric particulate matter source analysis model is: S, N, C, D, F, and R represent SO2 concentration, NO2 concentration, CO concentration, PM concentration, etc. 2.5 Concentration, PM 10 Characteristic parameters of concentration and humidity, when S=0, N=0, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When S=1, N=0, C=1, D=0, F=0, R=0, PM 10 Actual concentration>80μg / m 3 , PM 2.5 Actual concentration ≥100 μg / m 3 When , the source category of the corresponding feature matrix image is determined to be a coal-burning source; When S=0, N=0, C=0, D=0, F=0, and R=0, the source category of the corresponding characteristic matrix image is determined to be a partial secondary inorganic source; when S=1, N=0, C=1, D=1, F=1, and R=0, the actual concentration of SO2 is less than 20 μg / m 3 , PM 10 Actual concentration>80μg / m 3 When S=0, N=1, C=0, D=1, F=0, R=0, PM 10 Actual concentration ≥30μg / m 3 When S=1, N=0, C=0, D=0, F=0, R=0, the actual concentration of SO2 is less than 20μg / m 3 , PM 10 Actual concentration>80μg / m 3 When the source category of the corresponding feature matrix image is determined to be industrial source, in other cases, the source category of the corresponding feature matrix image is determined to be other source.

7. The method for source analysis of atmospheric particulate matter based on a convolutional neural network to construct a pollution matrix according to any one of claims 1 to 5, characterized in that: In step S4, the method for establishing the characteristic matrix images corresponding to different samples in the unknown pollution source includes the following steps: Step S4-1: SO2 concentration, NO2 concentration, CO concentration, PM 2.5 Concentration, PM 10 Concentration and humidity were fitted into a 6 × 6 matrix as pollution index parameters and used as a blank matrix; Step S4-2, normalizing the data of each pollution index parameter in different samples of the unknown pollution source during the pollution period to obtain characteristic parameters corresponding to each pollution index parameter in different samples of the known pollution source; Step S4-3: input the characteristic parameters corresponding to each pollution index parameter in different samples of known pollution sources into the blank matrix to obtain characteristic matrix images corresponding to different samples of unknown pollution sources.

8. The method for source analysis of atmospheric particulate matter based on a convolutional neural network to construct a pollution matrix according to any one of claims 1 to 5, characterized in that: Step S5 also includes the following processing: using an autoregressive model to determine the pollution period of the feature matrix image corresponding to different samples in the unknown pollution source, and drawing a pollution source change effect diagram corresponding to different pollution periods of the unknown pollution source within the pollution period.

9. The atmospheric particulate matter source analysis method based on convolutional neural network to construct pollution matrix according to claim 8 is characterized in that: According to the PM in the sample 2.5 The concentration determines the pollution period of the characteristic matrix images corresponding to different samples in the unknown pollution source, as follows: 2.5 Concentration <30 μg / m 3 When the characteristic matrix image of the corresponding sample is judged as the matrix of the non-polluting period; when 31μg / m 3 <PM 2.5 Concentration <70 μg / m 3 , the characteristic matrix image of the corresponding sample is determined as the matrix of the pollution coming period; when 71μg / m 3 <PM 2.5 Concentration < 500 μg / m 3 When PM 2.5 The concentration decreased for three consecutive values ​​and then recovered to PM 2.5 Concentration <30 μg / m 3 Before, the characteristic matrix image of the corresponding sample is determined to be the pollution decay period matrix.

Images